Abstract
Local climate zone (LCZ) has become a new standard classification scheme in urban landscapes and showed great potential in urban climate research. Traditional classifiers and ordinary neural networks only consider the spectral or local spatial features of the pixel, ignoring the effect of nonlocal information on the LCZ classification. The graph convolutional network (GCN) has been used to exploit the relationship between adjacent and global land covers owing to the ability to conduct flexible convolution over graphs. In this work, we integrated a convolutional neural network and two GCNs into an end-to-end hybrid framework and generated LCZs directly from the original images. Local-, regional-, and global-level features were extracted and grouped complementarily to foster better performance. Experiments were conducted in six cities around the world to verify the effectiveness of our method. Results showed that the average classification accuracy of the six cities reached 0.956 and performed better than any other comparable model. Ablation experiments also demonstrated the mutual promotion of the different modules. Finally, the small sample experiment provided a practical reference for the LCZ classification in the absence of samples in future.
What Problem Does This Paper Address?
Local Climate Zones (LCZs) are a standardized classification scheme used to understand urban heat islands and city-level climate impacts. They categorize landscapes into 17 distinct types based on physical structure (e.g., "compact high-rise") and land cover (e.g., "dense trees").
The primary challenge in mapping these zones is scale dependence. Traditional methods—such as Random Forests or standard Convolutional Neural Networks (CNNs)—primarily focus on local pixel data or small image patches. This often results in "fragmented" maps that ignore the broader spatial context. For instance, a park might be classified inconsistently depending on whether the analysis window is 200m or 1000m. To create accurate, consistent maps, a model must understand not just the individual pixel, but how that pixel relates to its immediate neighborhood and to similar zones located kilometers away.
Methods
The authors propose an end-to-end hybrid framework that extracts features at three distinct levels to capture the full context of urban environments:
- Local (ResNet): A standard residual network extracts fine-grained spectral and spatial features from the immediate vicinity of a pixel.
- Regional (RGCN): To ensure regional consistency and avoid "noisy" fragmented classifications, the authors use the SLIC algorithm to group pixels into "superpixels." These superpixels become nodes in a graph. The Regional GCN then learns the relationships between these adjacent blocks, effectively expanding the model's receptive field.
- Global (GGCN): Some LCZ types (like "compact mid-rise") appear in multiple, disconnected parts of a city. The Global GCN uses a downsampling strategy to build a lightweight, fully connected graph of the entire scene. This allows the model to capture "non-local" information—linking similar land types across the entire city to improve classification stability.
The features from these three branches are concatenated and passed through a Softmax classifier to produce the final LCZ map.
Datasets
The model was evaluated on six cities across three continents, providing a robust test of geographical diversity:
- Berlin, Rome, Paris: European cities with varying historical and modern layouts.
- Hong Kong: A high-density coastal metropolis.
- Sao Paulo: A sprawling South American city with high architectural variety.
- Wuhan: A multicentric Chinese city with complex water systems.
Data sources included Landsat-8 (9 bands at 100m resolution) and Sentinel-2 (for Wuhan). Reference ground-truth labels were sourced from the WUDAPT project and the 2017 IEEE GRSS Data Fusion Contest.
Key Results and Insights
- Superior Accuracy: The proposed method achieved an average Overall Accuracy (OA) of 95.6%, significantly outperforming the standard WUDAPT Random Forest (averaging ~74.5%) and other deep learning benchmarks like S5-CNN.
- Visual Consistency: Qualitatively, the resulting maps are much "cleaner." While pixel-based methods often produce a "salt-and-pepper" effect of misclassified pixels, the dual-GCN approach produces smooth, continuous zones that align better with actual urban geography.
- Dynamic Adjacency: Unlike standard GCNs where the graph is fixed, this model updates the "weights" of the edges between nodes during training based on feature similarity. This means the model learns which regions should influence each other.
- Small Sample Robustness: In experiments using as few as 5 to 25 samples per category, the hybrid model significantly outperformed traditional CNNs. This makes it highly practical for cities where extensive ground-truth labeling is unavailable.
Limitations and Future Work
While powerful, the model still faces challenges with highly imbalanced classes. For example, in Sao Paulo, the "Bare rock or paved" (LCZ-E) category had very few samples, leading to lower individual class accuracy despite high overall performance. Future research may explore multitemporal data to better distinguish vegetation types or advanced architectures like LSTMs to further refine spatial and temporal information weighting.
Summary updated